fix(eval): include ragged prompt tails in PE and unified evaluation

[Yuanjia-Lee]

Summary

_eval_pass in PE and UnifiedModel floored the eval set to whole chunks of
the training batch size (usable = n_prompts - n_prompts % chunk or n_prompts). The ragged tail was dropped before generation, so what got
scored depended on the training batch config: 13 prompts at batch size 8 are
scored as 8, at batch size 4 as 12. The default eval_num_prompts = batch_size is always divisible, which is why this never showed up.

On a 13-prompt set whose full-set winner is B but whose first 8 prompts favor
A, chunk=8 scores A=1.00 vs B=0.75 and picks the wrong model; the full set is
0.615 vs 0.846. Checkpoint selection via the #202 eval suites is affected.

The fix pads the tail chunk with replicated roots given fresh :eval-pad:
ids (the ARTrainer._pad_eval_inputs pattern, moved into eval_suites.py),
keeps every dispatch at the training batch size, verifies each batch's reward
count against roots x fan-out (so an engine-side mismatch fails loudly
instead of skewing the mean), and drops pad descendants, so the scored set is
the configured set. Below-chunk, divisible, and per-scorer mean semantics
are unchanged.

Related Issue

N/A

Test Plan

• Local harness (uncommitted, test: remove tests directory #99/test: remove tests directory #267) over real Sample types: with the
  fix every configured prompt is scored and the mean matches the per-sample
  oracle; divisible and below-chunk sets match legacy exactly.
• Real dispatch on one 4090: the patched _eval_pass drives a
  @distributed rollout+scorer role in a real worker; each root is scored
  exactly once and pad leaves stay out of the mean.
• ruff check / ruff format --check / check-docstring-lines /
  py_compile pass.

Compatibility / Risk

No config, checkpoint, data-format, or API changes. Eval numbers on
non-divisible sets change (that is the fix); divisible and below-chunk sets
are byte-identical.

Reviewer Notes

eval_suites.py gains pad_eval_inputs; pe.py and unified_model.py call
it in parallel in _eval_pass. ARTrainer._pad_eval_inputs is untouched (it
pads to the DP lcm, not the batch chunk). No open PR touches evaluation
batching.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.